Tencent Yuanbao Bazi: Text-Generated Destiny Symbols

Tencent Yuanbao is a general-purpose conversational model that generates text token by token. When it is asked to produce a bazi reading, the relations shown on the chart are written out rather than retrieved from a fixed reference system. This distinction matters for verification: a correct chart depends on fixed calendrical arithmetic, while a generated description of branch relations may include plausible-sounding items that do not correspond to any real pairing.

第三层核对：地支之间的关系

After the four pillars themselves have been checked, the next layer is the set of relations among the earthly branches. The branches are the twelve characters 子，丑，寅，卯，辰，巳，午，未，申，酉，戌, and 亥. In bazi analysis, relations such as combination, clash, punishment, and harm are said to exist between specific pairs of branches. These relations are not flexible or interpretive at the level of basic pairing. For any two branches, the relation, if any, is fixed.

The free charting tool on this site lists the combination, clash, punishment, and harm relations among the three columns when only the year, month, and day are supplied. Once the hour is added, the tool lists them across the four pillars. This produces a set of relations that can be checked one by one.

The key point about Tencent Yuanbao is that the chart relations are produced by a language model. A general chat model does not maintain a hidden bazi engine that looks up each pair against a fixed table. Instead, it writes the relations as part of its response. Because the model is generating text, it may include a relation that sounds familiar but does not exist in the standard correspondence, or it may omit a relation that should be present.

合冲刑害各自是固定对应

The six combinations, six clashes, mutual punishments, and mutual harms among the earthly branches are fixed correspondences. Any two branches can be checked directly against a table. The result does not change depending on who is doing the reading, what school is being used, or what the surrounding chart contains. If two branches form one of the six clashes, they form that clash. If they form one of the six combinations, they form that combination. The basic presence of the relation is not a matter of judgment.

The six clashes are formed by pairs of branches standing six positions apart. There are twelve branches in total, so this yields six pairs. The six combinations are also six fixed pairings. Punishment and harm have their own fixed correspondences, some of which involve more than a simple pair, but the underlying entries are stable.

Because these pairings are fixed, they can be used as a verification table. A person checking a generated chart does not need to know advanced theory. It is enough to have the list of six clashes, six combinations, the punishment groups, and the harm pairs. For each relation listed by the tool or model, the checker asks whether that exact relation name appears for that exact pair in the fixed table.

把工具列出的关系逐条对表

The checking method is mechanical. First, copy down every relation that the tool or model lists. For example, if the output says that 寅 and 申 form a clash, write that down. If it says that 子 and 未 form a harm, write that down. Do not skip any item because it looks plausible. Every listed relation is a claim about a fixed correspondence.

Second, take each copied item and compare it against the fixed table. If the table says that 寅 and 申 are one of the six clashes, the item matches. If the table does not list a harm relation between 子 and 未, the item is flagged as extra. If the table lists a relation between two branches that the output does not mention, that is a missing item.

Third, keep the extra items and the missing items separate. They indicate different kinds of failure. Do not mix them into a single error count, because they arise from different mechanisms and have different implications for the reliability of the output.

The fixed table is small. The six clashes are six entries. The six combinations are six entries. The punishment and harm lists are similarly limited. Checking every listed relation against this table usually takes only a few minutes. The total number of possible branch pairs in a four-pillar chart is small as well: four branches produce at most six two-branch combinations, so there are at most six possible pair relations to check, plus any additional grouped punishment entries.

多报与漏报说明的问题不同

Missing items and extra items point to different problems. A missing relation is often a sign of incomplete coverage. The model may have listed some relations but not all of them, perhaps because the response was cut short, because the model did not enumerate the full set, or because it attended only to the more familiar pairings. A missing item does not necessarily mean the model invented anything; it means the output was not exhaustive.

An extra item is more serious. If the output lists a relation that does not appear in the fixed correspondence, that relation was generated. It was written into the response even though no such relation exists between those branches. This is the signature of a language model producing text from patterns of usage rather than from a table lookup. An extra relation may look structurally similar to a real one, and it may use the same vocabulary, but it has no basis in the fixed pairings.

For example, a model might write that two branches form a combination even when they are not one of the six fixed combination pairs. The statement may be fluent and confident, but the verification table shows it is not a genuine entry. The distinction is not about interpretive nuance; it is about whether the pairing exists at all in the fixed correspondence.

The practical consequence is that a user should treat extra relations as errors of generation, not as alternative readings. If the same model produces different relation lists on different runs with the same four pillars, that instability is further evidence that the relations are being generated rather than looked up.

这一层为什么比四柱更容易松动

This layer fails differently from the four-pillar layer. When the four pillars themselves are wrong, the error is a calendrical or arithmetic error. The stem and branch for a given date and time can be computed from the sexagenary cycle and the solar terms, and a mistake there means the calculation went wrong. The four pillars are either correct or incorrect, and the correct result is the same regardless of the source.

The relation layer can be wrong even when the four pillars are correct. A model may compute or reproduce the four pillars accurately, then fill in the relations from memory of common patterns. Because the relations are written as prose, the model can produce a fluent list that includes both real and fabricated items. The four pillars may be exactly right, while one of the listed relations is not in the fixed table. This is a failure of description, not of calculation.

This is why the relation layer is more likely to loosen under repeated use. A deterministic bazi engine will produce the same relation list every time for the same chart, because it is checking fixed pairs. A language model may produce different wording, different coverage, and different extra items on different runs. The underlying chart may remain stable while the relation list changes. That instability is not proof that the chart is wrong; it is evidence that the relation text was generated.

The verification routine therefore treats the two layers separately. First confirm the four pillars against a known calendar or a trusted charting tool. Then check each relation against the fixed correspondence. The first check catches calculation errors. The second catches generated content that has no fixed basis. Because the relation table is small and the number of branch pairs is limited, the second check is quick. The main requirement is to keep the fixed table at hand and to resist the assumption that a fluent relation list is necessarily a correct one.